Requirements with AI

Requirements work is translation: fuzzy stakeholder wants → precise statements a builder can build and a tester can test. AI speeds up the drafting dramatically — but drafting was never the hard part. The hard part is finding what's missing, ambiguous, or contradictory before it becomes expensive, and that's where AI-as-adversary earns its seat.

Drafting: fine, fast, unremarkable

The adversarial passes (this is the value)

• Ambiguity hunt: 'Review these requirements as a hostile contractor looking for ambiguity to exploit. For each requirement, what are two different things it could mean? Which words are doing vague work?' — Words like timely, accurate, by store, and refund (gross? net of re-shipments? cancellations included?) all crack open under this pass. Every ambiguity you resolve now is a meeting you don't have later.
• Edge-case generation: 'List 15 edge cases these requirements don't address' — partial refunds, refunds spanning two reporting months, a refund for an order placed in a previous system, store credit vs. card refund. You'll keep six, discard nine, and be glad you saw all fifteen.
• The misreading stakeholder: 'Read this as a busy support manager who skimmed it. What will they THINK they're getting that they're not?' — Expectation gaps found on paper cost nothing; found at delivery, they cost the relationship.
• The traceability check: 'For each requirement, which interview theme or business goal does it trace to? List any requirement that traces to nothing.' — Orphan requirements are either missing discovery or invented scope; both need a decision, not a shrug.

Notice the pattern across all four passes: AI generates candidates; you adjudicate. The model is spectacular at 'what could be wrong here' breadth and mediocre at knowing which of them matters at your company. That division of labor — machine breadth, human judgment — is the course's recurring shape, and it's why the analyst stays in the loop no matter how good the model gets.